Applied Machine Learning for Senior Leaders

Today…

Solve a prediction challenge
Explore and understand our data with Python
Think about how our data is connected, and test our theories
Use our findings to build our first model, and see how it performs
Use our model to make predictions

Tutorial: bit.ly/ml-for-board

Our Data, and our Challenge

We’ve built a fake dataset
Each row represents a benefit claimant, that has been suspected of fraud, over the last 12 months
Divided into:
- a training dataset (270 individuals) where a human has manually verified whether fraud was commited
- a test dataset (30 indivuals) where we’ll aim to predict whether or not fraud was committed
Can we build a model to predict whether claimants have comitted fraud?

Relationships in our data - Correlation

To make a model, we need to understand how one variable in our data is affected by the others
A good place to start is with correlation
Correlation is a statistical measure of how much two things are related to each other, whether they increase or decrease together, or if there is no relationship at all.

start with slides run through
Remember what we’re trying to achieve here: understanding how different parts of our data connect, so we can build a model
Correlation might sound complicated, but it’s really just a mental shortcut to understand how to parts of our data relate
It starts at 1 (perfect correlation): for example, if your weight went up by 1 kilo for every 10 cms you grew
In the real world, we rarely have correlations of 1 - but we might have correlations between 0.3-0.5, described as moderate
And of course, lots of different data points have no correlation, around 0, meaning no apparent connection
And you can also have negative correlation - that’s just flipping the lines the other way around
So remember our first example? A correlation of -1 would mean mean you lost 1 kilo for every 10 cms you grew

Examples of Correlation

How well is a person’s age in years correlated with…

their age in months

their income

the length of their name

Let’s code it!

training[['age','years_in_labour_force']].corr()

	age	Years in labour force
age	1.0	0.749

How would you interpret these coefficients?
Can you find features linked to fraud?

Building our first regression

model = smf.ols(formula='years_in_labour_force ~ age', data=training)
results = model.fit()
results.summary()

For now, we’ll focus on our R-squared value

Understanding R-Squared

R-squared is a measure that tells us how well our data fits our model
I like thinking of it as how “good” our model is, in percentage terms
So a model with a R-squared of 1.0 fits 100% of our data, while 0.5 would fit 50%
It’s a little like correlation…but with a lot more potential!

Regression Example - Restaurant Bills

model = smf.ols(formula='tips ~ bill', data=all_bills)

model = smf.ols(formula='tips ~ bill + gender', data=all_bills)

How does it change when we add features?

Making our predictions

Now we have our model, how do we make predictions?
We just pass it our test data!

model = smf.ols(formula='committed ~ age + years_in_labour_force + years_of_further_education+', data=train)
results = model.fit()
predictions = results.predict(test)

	committed	prediction
person_id
203	1	0.84
266	1	0.58

	age	committed	prediction
person_id
42	29.50843	1	True
281	50.42585	1	True
295	29.02093	0	True
157	34.13484	1	True
238	33.00470	1	True

From here… to AI!

We’ve understood how our data connects, and combined it into a model
The same core methods and principles apply to both what you’ve done today, and cutting edge AI

Thank you!

avarotsis@no10.gov.uk